Profession Calculators
Tech & ITPopular

Vector Database & Embedding Storage Sizer

Size vector database disk storage and planning RAM from current embedding dimensions, quantization, metadata, replicas, and index type, with verified Pinecone, Qdrant, and Zilliz pricing inputs.

Share:
Vector Index & Dataset Parameters

Standard 1,536-dim embedding model for general RAG search.

Infrastructure Sizing Projections
Configure your vector dimension and document count, then click Calculate to generate memory and disk allocations.

Embed This Calculator on Your Website

Add this free calculator to your blog, website, or CMS with a simple copy-paste embed code.

Introduction

Vector storage depends on math, but managed database pricing depends on provider-specific units that cannot be inferred from QPS alone. The core storage calculation remains dimensions multiplied by bytes per dimension multiplied by vector count. Current model choices have also changed. Google Gemini embedding documentation describes Gemini Embedding 2 with a default 3,072-dimensional output that can be truncated to smaller sizes. OpenAI embedding documentation still lists text-embedding-3-small and text-embedding-3-large, with the large model using 3,072 dimensions. On the database side, Pinecone pricing lists $0.33 per GB per month for Standard on-demand storage, $16 to $18 per million read units, and $4 to $4.50 per million write units depending on cloud and region. Qdrant pricing publishes a free 1 GB RAM, 4 GB disk tier but directs production users to its calculator. Zilliz list pricing starts dedicated serving compute at $0.273 per CU-hour in the referenced AWS region.

What This Calculator Does

Choose a current embedding model or enter a custom dimension count. Add vector count, metadata size, numeric precision, index type, replica count, and target QPS. The calculator estimates raw vector bytes, replicated metadata, index overhead, total disk storage, and a planning RAM target. It does not invent a monthly provider bill from QPS. Query price depends on topK, vector dimensions, filters, namespaces, read-unit formulas, capacity-unit sizing, region, and service plan. The provider cards therefore show current published pricing inputs and direct you to official calculators for a quote.

The Formula

Raw Vector Bytes = Dimensions x Bytes per Dimension x Vector Count x Replicas | Metadata Bytes = Metadata Bytes per Vector x Vector Count x Replicas | Estimated Index Overhead = Raw Vector Bytes x Index Overhead Ratio | Total Disk = Raw Vector Bytes + Metadata Bytes + Index Overhead | Planning RAM = Resident Vector and Index Data x Working Buffer

Float32 uses 4 bytes per dimension, Float16 uses 2, INT8 uses 1, and a binary vector uses one bit per dimension. A 3,072-dimensional Float32 vector therefore occupies 12,288 bytes before row, graph, metadata, allocator, and database overhead. HNSW overhead varies with M, graph layers, identifiers, implementation, and allocator behavior. The calculator uses a planning ratio, not an exact byte count. Replicas multiply stored data. QPS is displayed for planning but is not converted directly into provider read units because each provider measures work differently.

Step-by-Step Example

1

Select an embedding output size

Choose Gemini Embedding 2 at its default 3,072 dimensions, or enter the truncated dimension count actually configured in your API request.

2

Enter vectors, precision, and replicas

Enter 2,000,000 vectors, Float32, and two replicas. Raw vector storage alone is approximately 45.8 GiB before metadata and index overhead.

3

Choose the index architecture

Select HNSW for a high-recall graph estimate, or select IVF-PQ when your actual database uses compressed product quantization.

4

Price with provider-specific units

Take the disk and RAM result to Pinecone, Qdrant, Zilliz, or your cloud instance calculator. For Pinecone, add storage, measured read units, measured write units, egress, and the applicable plan minimum.

Real-World Use Cases

Gemini Embedding 2 migration

A search team moving from 768 dimensions to a 3,072-dimension default measures the 4x raw storage impact before deciding whether to truncate output.

OpenAI embedding index planning

A RAG service compares text-embedding-3-small at 1,536 dimensions with text-embedding-3-large at 3,072 dimensions under the same precision and document count.

Quantization review

An engineer checks the theoretical storage reduction from Float32 to INT8 before running retrieval-quality benchmarks on the database's actual quantizer.

Managed versus self-hosted sizing

A platform team uses the same RAM and disk workload to request managed quotes and to price pgvector or Qdrant on current regional instances.

Comparison

Current inputPublished detailWhat still needs measurement
Gemini Embedding 23,072 dimensions by default; configurable truncationChosen dimensions and corpus quality
OpenAI text-embedding-3-small1,536 dimensionsChosen dimensions if shortened
OpenAI text-embedding-3-large3,072 dimensionsChosen dimensions if shortened
Pinecone Standard On-Demand$0.33/GB-month; $16 to $18 per 1M reads; $4 to $4.50 per 1M writesActual read and write units, egress, region
Qdrant CloudFree tier: 1 GB RAM and 4 GB diskProduction cluster price from official calculator
Zilliz Cloud$0.273/CU-hour dedicated; $0.41/CU-hour query or indexing in referenced AWS regionRequired CUs, storage, transfer, region

Common Mistakes to Avoid

  • Turning QPS directly into dollars. TopK, dimensions, filter complexity, cache behavior, namespaces, and provider unit formulas affect query cost.

  • Treating HNSW overhead as a fixed universal percentage. M, efConstruction, identifiers, replicas, and implementation details change graph size.

  • Using the model's maximum dimension instead of the configured output dimension. Gemini Embedding 2 and some other models permit smaller outputs.

  • Counting raw vectors but not metadata or replicas. Filter payloads and high-availability copies can exceed vector storage.

  • Publishing a managed monthly price without a provider calculator. Qdrant, Zilliz, Pinecone, and cloud regions use different billing dimensions and minimums.

Frequently Asked Questions

Is 3,072 dimensions current in August 2026?

Yes. Gemini Embedding 2 defaults to 3,072 dimensions, and OpenAI text-embedding-3-large is also 3,072 dimensions. Both should be sized using the output dimension you actually store.

Why was the old provider cost estimate removed?

It converted QPS into invented read-unit and RAM prices. That was not reliable. The revised calculator reports technical capacity and current published pricing inputs without pretending they produce a universal quote.

Does INT8 always preserve retrieval quality?

No. It reduces theoretical storage but recall impact depends on model distribution, quantizer, corpus, and rescoring. Benchmark before migration.

How should I use target QPS?

Use it to design a load test with real topK, filters, concurrency, and latency targets. Then use measured provider units or required cluster capacity for pricing.

Are GB and GiB interchangeable?

No. Provider billing often uses decimal GB, while memory tools commonly report binary GiB. This calculator converts bytes using 1,073,741,824 bytes per GiB.

Accuracy and Disclaimer

Model dimensions and provider pricing were researched from official Google, OpenAI, Pinecone, Qdrant, and Zilliz pages and verified August 19, 2026. Index overhead and RAM remain planning estimates. Provider charges vary by plan, cloud, region, workload shape, transfer, minimum spend, and negotiated discount. Confirm final pricing with official calculators and production benchmarks.

Conclusion

Use this result to choose a realistic benchmark environment, not as a replacement for load testing. Generate representative vectors, filters, and topK requests, then measure recall, p95 latency, memory, and provider usage units. For the generation layer, use the LLM API Token Cost Estimator. For traffic between your application and managed database region, use the Cloud Egress and Data Transfer Cost Estimator.